CN107358142A - Polarimetric SAR Image semisupervised classification method based on random forest composition - Google Patents
Polarimetric SAR Image semisupervised classification method based on random forest composition Download PDFInfo
- Publication number
- CN107358142A CN107358142A CN201710341085.3A CN201710341085A CN107358142A CN 107358142 A CN107358142 A CN 107358142A CN 201710341085 A CN201710341085 A CN 201710341085A CN 107358142 A CN107358142 A CN 107358142A
- Authority
- CN
- China
- Prior art keywords
- mrow
- data
- random forest
- msub
- similarity relation
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/10—Terrestrial scenes
- G06V20/13—Satellite images
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
- G06F18/243—Classification techniques relating to the number of classes
- G06F18/24323—Tree-organised classifiers
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Evolutionary Biology (AREA)
- Evolutionary Computation (AREA)
- Bioinformatics & Computational Biology (AREA)
- General Engineering & Computer Science (AREA)
- Artificial Intelligence (AREA)
- Life Sciences & Earth Sciences (AREA)
- Astronomy & Astrophysics (AREA)
- Remote Sensing (AREA)
- Multimedia (AREA)
- Image Analysis (AREA)
Abstract
A kind of semi-supervised Classification of Polarimetric SAR Image method based on random forest composition of disclosure of the invention.Mainly solves the problem of deficiency of similarity relation between expression sample point in existing sorting technique, non-utilization space information.Its step is:Input Polarimetric SAR Image initial data;The correlated characteristic of data is extracted, obtains data set;Build initial random forest model;Two graders are trained using two different attribute sample sets, for assisting to train semi-supervised Random Forest model;Optimize semi-supervised Random Forest model;Build similarity relation figure;Build spatial information figure;Merge similarity relation figure and spatial information figure obtains the similarity relation matrix between sample point;Image is classified and calculates accuracy.The present invention constructs more affine similarity relation figure and spatial information using semi-supervised random forests algorithm, improves the classification accuracy rate of Polarimetric SAR Image.For the civil and military such as geological prospecting, rescue and relief work, target identification field.
Description
Technical field
The invention belongs to technical field of image processing, relates generally to Classification of Polarimetric SAR Image, is specifically a kind of be based at random
The Polarimetric SAR Image semisupervised classification method of forest composition, available for the terrain classification and target identification to Polarimetric SAR Image.
Background technology
Polarization SAR is that one kind is to utilize synthetic aperture principle, high-resolution microwave imaging radar is realized, not only with whole day
Time, round-the-clock, high resolution, can side view imaging the advantages that, while SAR images have abundant detailed information, important texture
Feature and obvious atural object geometry, it can be widely applied to the numerous areas such as military, agricultural, navigation, geographical monitoring.In the world
Remote sensing fields are highly valued, therefore Classification of Polarimetric SAR Image has turned into an important research side of polarization SAR information processing
To.
The purpose of Classification of Polarimetric SAR Image is the polarization measurement data obtained using airborne or borne polarization SAR sensor,
According to the property of pixel, it is determined that the classification belonging to each pixel.It is an important content of image interpretation technology, is it
The basis that it is applied.As application of the polarization SAR in life is military is more and more extensive, for the essence of Classification of Polarimetric SAR Image
Exactness requires also more and more higher, and in the sorting technique of some existing Polarimetric SAR Images, some is only just with label
The information of data is classified, such as the sorting technique supervised, such as KNN methods.In order to be combined with label data and nothing
The information of label data is classified, and has scholar to propose semi-supervised sorting technique, application is more to be included based on figure
Semi-supervised method, such as the semisupervised classification method based on anchor point figure, this method by between calculating image pattern point away from
From build anchor point figure matrix, however polarization SAR data be it is nonlinear, it is simple calculate between image pattern point away from
It is inaccurate from come the similarity relation that represents between sample point.The above method is all first by having exemplar and unlabeled exemplars
A graph model is built as summit, all assigns a weights to the side between each opposite vertexes, weights represent two samples
Similarity between point.Afterwards, it is that the class label of marked sample is passed to nothing by grader by side according to majorized function
Marker samples, so as to classify to unmarked sample.In core low-rank figure, each sample point in data set is found out first
K Neighbor Points, then calculate the partial reconstruction weight matrix of sample point.In anchor point figure, first by clustering method either
Randomly select a part of sample point and form an anchor point collection, then calculate the similar diagram matrix between anchor point and original sample, surpass
Figure is then to consider several data points together, and they are connected together with a line.
In the structure of above-mentioned graph model, sample point similarity relation is strong and weak typically by two data of calculating in image
The distance between point represents, but the method for distance to a certain extent can not be very accurate between this simple calculating data
Structural relation between true SAR data of the expression with nonlinear organization, and not between consideration image pattern point
Spatial information, cause the effect of classification poor.
The content of the invention
It is an object of the invention to the deficiency for above-mentioned prior art, propose a kind of classification accuracy rate it is higher based on
The Polarimetric SAR Image semisupervised classification method of machine forest composition, the Random Forest model of supervision is improved to semi-supervised random gloomy
Woods model, the use of label is greatly reduced, utilize the random forest method for being suitable for that there are non-linear behavior polarization SAR data
It is patterned, to strengthen the compatibility of figure, the spatial information of image pixel sample point has been it is possible to additionally incorporate, so as to improve image
Classification accuracy rate.
1st, a kind of Polarimetric SAR Image semisupervised classification method based on random forest composition, it is characterised in that including just like
Lower step:
(1) input picture:The initial data of Polarimetric SAR Image is obtained from polarimetric SAR image data.
(2) feature extraction:To the polarization SAR initial data of input, it is decomposed, extract data relevant T features,
Cloud features, Freeman features, Span features, obtain the data set X of 15 dimensions altogether.
(3) initial random forest is built:Data set X is upset at random, builds property set respectively, wherein first property set
For X1, comprising relevant T features, Cloud features, Span features, second property set is X2, include relevant T features, Freeman
Feature;It is F to build initial forest0, the sample per class 1% is chosen as exemplar collection, is designated as Xl, remaining is without label sample
This collection, it is designated as Xu, the loss of exemplar and the loss of unlabeled exemplars are combined, unlabeled exemplars and exemplar are used
Same loss function builds initial random forest model F0。
(4) semi-supervised Random Forest model is trained:Initialization training iterations S=0, s mark is chosen from exemplar
Two graders are originally respectively trained in signed-off sample, with first property set X1In s exemplar train first KNN graders f1,
With second property set X2In s exemplar train second KNN graders f2, for each unmarked sample, if two
The result that individual grader is classified to it is consistent, then the unmarked sample is the high sample of confidence level, takes all confidence levels
High unlabeled exemplars, a respective label value is given, adds it to tally set XlIn, exemplar collection is updated, then will
Marker samples collection after renewal is as training set to random woods model FsCarry out semi-supervised training.
(5) semi-supervised Random Forest model is optimized:Using definitive operation process (Deterministic Annealing
Process, DA)
To carry out optimization processing, by introducing a category distribution probability without label dataWill be without label data
Part be added in optimization aim;Give the data false segmentation rate initial value of a block mold:To control optimization,
The outer data false segmentation rate of whole bag is calculated after carrying out a suboptimizationWhenWhen, stop optimization, otherwise carry out next suboptimum
Change, go to step (4), renewal annealing temperature Ts+1=0.9Ts, training iterations S=S+1;Until obtain optimal random
Forest model.
(6) similarity relation figure is built with Random Forest model:Using the Random Forest model trained to the mark in image
Signed-off sample sheet and unlabeled exemplars carry out building similarity relation figure, obtain a similarity relation figure matrix W;It is non-using being suitable for having
The random forest method of linear feature polarization SAR data is patterned, therefore obtained similarity relation figure matrix W has preferably
Compatibility.
(7) spatial information figure is built:For each pixel samples point in image, it and four pictures of surrounding all around
Plain sample point similarity relation value is 1, and then value is 0 with other pixel samples point similarity relations, builds spatial information figure matrix
G。
(8) similarity relation figure and spatial information figure are merged:Similar diagram matrix W and spatial information figure matrix G are combined and drawn
The final figure matrix Z=W+ λ G, wherein λ ∈ (0,1) that represent similarity relation between pixel samples point in image.
(9) Polarimetric SAR Image is classified and calculates classification accuracy rate:The similarity relation figure matrix Z that will have been obtained
Polarimetric SAR Image is classified in semisupervised classification method for scheming holding, obtains the sorted classification of each pixel
Label matrix Y;Each pixel is coloured according to class label matrix Y, the image after output category, and calculate classification
Accuracy.
The present invention technical thought be:The Random Forest model of supervision is improved to reference to two KNN graders semi-supervised
Random Forest model, it is optimized using definitive operation method, obtains optimal semi-supervised Random Forest model, profit
Image pixel sample point is carried out with semi-supervised Random Forest model to build similarity relation figure matrix W, then utilizes image slices
The spatial information structure spatial information figure matrix G of plain sample point, merges similarity relation figure matrix W and spatial information figure matrix G is obtained
To final similarity relation figure matrix Z, finally similarity relation figure matrix is used in figure keeping method enter Polarimetric SAR Image
Row classification.
Compared with prior art, it is an advantage of the invention that:
1, the random forest method of supervision is improved to semi-supervised random gloomy by the present invention by using two KNN graders
Woods method.On the one hand reduce the use of exemplar, and sufficiently make use of the information of unlabeled exemplars, to a certain degree
On improve the classification accuracy rate of Polarimetric SAR Image.
2, when being patterned, the present invention utilizes random forests algorithm, is determined by movement images sample point in random forest
Whether whether classification path in plan tree identical similar between sample point to judge, similar between image pattern point so as to construct
Graph of a relation.The similarity relation between sample point is represented compared to using the distance between image pattern point, the inventive method is more
Adapt to the non-linear behavior of polarization SAR data.Therefore the similarity relation figure of structure is more affine, contributes to Polarimetric SAR Image
Classification.
3, the present invention considers the spatial information between image pattern point, utilizes the spatial neighbors relation structure between sample point
A spatial information figure is produced, with reference to the image pixel sample similarity relation figure constructed in 2, is further improving the classification of image just
True rate.
Brief description of the drawings
The implementation process figure of Fig. 1 present invention;
The experimental result picture of the existing methods of Fig. 2 and the inventive method, Fig. 2 (a) are that the inventive method and existing contrast are calculated
Method is used for the original graph tested, i.e. Pauli figures;Fig. 2 (b) is standard drawing corresponding to original graph, and emulation Polarimetric SAR Image mark
Label figure;Wherein Fig. 2 (c) is the result figure classified by KNN methods to Fig. 2 (a);Fig. 2 (d) is with Hyper methods to Fig. 2 (a)
The result figure classified;Fig. 2 (e) is the result figure classified using SSA methods to Fig. 2 (a);Fig. 2 (f) is the present invention
The result figure that method is classified to Fig. 2 (a).
Embodiment
Below in conjunction with the accompanying drawings to the detailed description of the invention
Embodiment 1
Because of the development of remote sensing technology, widely should have in fields such as environmental monitoring, earth resources survey, military systems
With the demand handled Polarimetric SAR Image also continues to increase, some existing polarization SAR sorting techniques, includes point of supervision
Class method, classification accuracy rate is relatively low, also have some semi-supervised methods based on figure, be mostly by calculate image pattern point it
Between distance build similarity relation figure matrix, but polarization SAR data are nonlinear, the similar passes of this patterning process
System is inaccurate.Therefore, the present invention proposes a kind of Polarimetric SAR Image semisupervised classification method based on random forest composition, ginseng
See Fig. 1, comprise the following steps:
(1) input picture:The initial data of Polarimetric SAR Image is obtained from polarimetric SAR image data file.If
In practical application, first image is read in using PolSAR softwares, each pixel in image is then processed into data file,
The primitive character of image is stored in the form of a file.
(2) feature extraction:To the polarization SAR initial data of input, it is decomposed, extract data relevant T features,
Cloud features, Freeman features, Span features, obtain the data set X of 15 dimensions altogether.
(3) initial random forest is built:Data set X is upset at random, builds property set respectively, wherein first property set
For X1, comprising relevant T features, Cloud features, Span features, second property set is X2, include relevant T features, Freeman
Feature;It is F to build initial forest0, the sample per class 1% is chosen as exemplar collection, is designated as Xl, remaining is without label sample
This collection, it is designated as Xu, the loss of exemplar and the loss of unlabeled exemplars are combined, unlabeled exemplars and exemplar are used
Same loss function builds initial random forest F0。
(4) semi-supervised random forest is trained:Initialization training iterations S=0, s label sample is chosen from exemplar
Originally two graders are respectively trained, with first property set X1In s exemplar train first KNN graders f1, with
Two property set X2In s exemplar train second KNN graders f2, and it is random gloomy using the assistance of the two graders
Woods carries out semi-supervised training, unmarked sample is classified by the synergy of two KNN graders first, for each
Individual unmarked sample, if the result that two graders are classified to it is consistent, the unmarked sample is the high sample of confidence level
This, takes the high unlabeled exemplars of all confidence levels, gives a respective label value, add it to tally set XlIn, renewal
Exemplar collection, then using the marker samples collection after renewal as training set to random woods FsCarry out semi-supervised training.
(5) semi-supervised random forest is optimized:Using definitive operation process (Deterministic Annealing
Process, DA) optimization processing is carried out, pass through and introduce category distribution probability without label dataWill be without number of tags
According to part be added in optimization aim;Give the data false segmentation rate initial value of a block mold:It is excellent to control
Change, the outer data false segmentation rate of whole bag is calculated after carrying out a suboptimizationWhenWhen, stop optimization, it is believed that obtained
Optimal Random Forest model.Otherwise next suboptimization is carried out, goes to step (4), renewal annealing temperature Ts+1=0.9Ts, instruction
Practice iterations S=S+1;Until having obtained optimal Random Forest model.
(6) random forest structure similarity relation figure:Using the random forest trained to the exemplar and nothing in image
Exemplar carries out building similarity relation figure, obtains a similarity relation figure matrix W;Having obtained similarity relation figure matrix W must
Similarity relation figure is arrived.The present invention utilizes the random forest method progress structure for being suitable for having non-linear behavior polarization SAR data
Figure, therefore obtained similarity relation figure matrix W has more preferable compatibility.
(7) spatial information figure is built:For each pixel samples point in image, it and four pictures of surrounding all around
Plain sample point similarity relation value is 1, and then value is 0 with other pixel samples point similarity relations, whole so as to obtain an expression
The figure matrix of spatial relation between individual image pattern point, is designated as spatial information figure matrix G.
(8) similarity relation figure and spatial information figure are merged:By the sky in the similar diagram matrix W in step (6) and step (7)
Between hum pattern matrix G be combined and draw the final figure matrix Z=W+ λ G for representing similarity relation between pixel samples point in image,
Wherein λ ∈ (0,1).
(9) Polarimetric SAR Image is classified and calculates classification accuracy rate:The similarity relation figure matrix Z that will have been obtained
Polarimetric SAR Image is classified in semisupervised classification method for scheming holding, obtains the sorted classification of each pixel
Label matrix Y;Each pixel is coloured according to class label matrix Y, the image after output category, and calculate classification
Accuracy.
It is more and more extensive with the application of polarization SAR, for Classification of Polarimetric SAR Image precise requirements also increasingly
Height, in the sorting technique of some existing Polarimetric SAR Images, such as the sorting technique supervised, poor effect, in order to combine profit
To be classified with label data with information without label data, some are suggested based on semi-supervised sorting technique is schemed, these
Method carries out building anchor point figure matrix by calculating the distance between image pattern point, but polarization SAR data are non-linear
, the distance between simple calculating image pattern point is inaccurate come the similarity relation represented between sample point.The present invention passes through
Using the method for improved semi-supervised random forest, divided using image pattern point self attributes in random forest decision tree
Path come instruct build image pixel sample point between similarity relation figure matrix, the similarity relation figure matrix so obtained is more
Meet the nonlinear characteristic of polarization SAR data.So as to improve the classification accuracy rate of Polarimetric SAR Image.
Embodiment 2
Based on the Polarimetric SAR Image semisupervised classification method of random forest composition with embodiment 1, in step (4) of the present invention
Train the process of semi-supervised random forest as follows:
4a, initialization training iterations S=0, choose s exemplar from exemplar and two classification are respectively trained
Device, s value is depending on the size of image actual size, s=80 in the present embodiment;
4b, with the first property set X180 exemplars of middle selection train first KNN graders f1, with the second attribute
Collect X280 exemplars of middle selection train second KNN graders f2;
4c, the semi-supervised training of Random Forest model progress is assisted using the two graders.Assisted using two graders
It is in order to ensure the efficiency of training, relative to using more efficient for classifier training that Random Forest model, which is trained,
Accurately.Unmarked sample is classified by the synergy of two KNN graders first, for each unmarked sample
This, if the result that two graders are classified to it is consistent, illustrates that the possibility that the unmarked sample is certain class is very big, then
The unmarked sample is the high sample of confidence level, takes the high unlabeled exemplars of all confidence levels, gives a respective label value,
Add it to tally set XlIn, update exemplar collection, then using the marker samples collection after renewal as training set to random
Woods model FsIt is trained.Obtain new Random Forest model Fs+1。
In the present invention, the exemplar after renewal is concentrated, and both comprising the sample for being originally used for label, is also included and is originally used for
Sample without label, therefore be semi-supervised Random Forest model using the Random Forest model of the tally set training after renewal.
Embodiment 3
Based on the Polarimetric SAR Image semisupervised classification method of random forest composition with embodiment 1-2, step (5) is middle to be optimized
The process of semi-supervised random forest is as follows:
5a, the present invention are carried out using definitive operation process (Deterministic Annealing process, DA)
Optimization processing, by introducing a category distribution probability without label dataWhereinAdjusted for normalization.By without label data
Part is added in optimization aim, and expression formula is as follows:
Sum term Part ITo have label data loss, Section 2For the expectation lost without label data, Section 3It is expressed as no label
The comentropy of data distribution;α is expectation weighted value and α ∈ [0,1] without label data loss, and value is in the present embodiment
0.5, T for annealing temperature variable, in the present embodiment initial value be set as 1, when T is 0, the formula turns to initial random gloomy
Woods model loss function;
5b, the data false segmentation rate initial value for giving a block mold:To control optimization, a suboptimization is carried out
After calculate the outer data false segmentation rate of whole bagWhenWhen, stop optimization, it is believed that have been obtained for optimal random forest
Model.Otherwise next suboptimization is carried out, goes to step (4), renewal annealing temperature Ts+1=0.9Ts, training iterations S=S+
1;Until having obtained optimal Random Forest model Fs+1。
The present invention optimizes by using definitive operation mode to semi-supervised Random Forest model, in the model,
The loss of label data is not only allowed for, is also lost the prediction category without label data as optimization aim, and consider
Predict the comentropy of category probability distribution so that expectation is not only ensure that in optimization, and controls variance, is maximised all
The interval of sample.
Embodiment 4
Based on the Polarimetric SAR Image semisupervised classification method of random forest composition with embodiment 1-3, step (6) is middle to be optimized
The process of semi-supervised random forest is as follows:
6a, in Random Forest model, the number of plies of each decision tree is identical, it is assumed that the number of plies t, a pair of data point (xi,
xj) from root node γ, divide by feature layer by layer, most a pair of data point (x of Zhongdaoi,xj) belonging to child node liAnd lj, a pair
Two data point x in data pointiAnd xjThe path of process is expressed as:
QiRepresent data point xiPath, QjRepresent data point xjPath, γ be decision tree root node.For number
Strong point xiBy the decision tree internal node on path;For data point xjBy the decision tree internal node on path.
6b, for each to data point (xi,xj), with the similarity relation represented between them, the similarity relation of all data
A similar diagram matrix W is formed, similarity relation figure matrix W has been obtained, has just obtained similarity relation figure, then had:
wijIt is data point xiWith data point xjBetween similarity relation value.
The present invention is using being suitable for that there is the random forest method of non-linear behavior polarization SAR data to be patterned, therefore
Obtained similarity relation figure matrix W has more preferable compatibility.
A more detailed example is given below, the present invention is further described
Embodiment 5
Semi-supervised Classification of Polarimetric SAR Image method based on random forest composition is the same as embodiment 1-4, reference picture 1, the present invention
Specific implementation step it is as follows:
Step 1: input picture, the initial data of Polarimetric SAR Image is obtained from polarimetric SAR image data file, is joined
See Fig. 2 (a), label matrix L obtained according to the atural object distributed intelligence of Polarimetric SAR Image, referring to Fig. 2 (b), Fig. 2 (b) be exactly by
The image that label matrix L is directly generated, different color lumps represents different atural object in image, and same atural object is distributed in label
Represented in matrix by same category label, as shown in the legend of Fig. 2 bottoms, such as the 1st category label corresponding to red block
For 1, the 2nd category label corresponding to green block is 2 ..., and category label corresponding to last blue block is 9.
If in actual applications, first image is read in using the PolSAR softwares for handling Polarimetric SAR Image, then figure
Each pixel as in is processed into the form of data, the initial data of image is stored in the form of a file, as pole
Change the initial data of SAR image.
This example uses polarization SAR geo-objects simulation image, and size is 120 × 150, and the emulation data have 18000 samples,
Each sample corresponds to a pixel on Polarimetric SAR Image, referring to Fig. 2 (a).
Step 2: feature extraction, to the polarization SAR initial data of input, it is decomposed, and the relevant T for extracting data is special
Sign, Cloud features, Freeman features, Span features, obtain the data set of 15 dimensions altogetherN is sample in image
This total number, xiRepresent i-th of sample.
Step 3: build initial random forest model:Data set X is upset at random, builds property set respectively, wherein first
Individual property set is X1, comprising relevant T features, Cloud features, Span features, second property set is X2, comprising relevant T features,
Freeman features;It is F to build initial forest0, the sample per class 1% is chosen as exemplar collection, is designated as Xl, remaining is nothing
Exemplar collection, is designated as Xu, the loss of exemplar and the loss of unlabeled exemplars are combined, to unlabeled exemplars and label
Sample builds initial random forest F using same loss function0。
Step 4: train semi-supervised Random Forest model:
4a, initialization training iterations S=0, choose s exemplar from exemplar and two classification are respectively trained
Device, s value is depending on the size of image actual size, s=100 in the present embodiment;
4b, with the first property set X1100 exemplars of middle selection train first KNN graders f1, with the second attribute
Collect X2100 exemplars of middle selection train second KNN graders f2, obtained two graders trained.
4c, the grader trained using the two assist Random Forest model to carry out semi-supervised training:Utilize two points
It is in order to ensure the efficiency of training, for using a classifier training that class device, which assists Random Forest model to be trained,
More efficiently and accurately.Unmarked sample is classified by the synergy of two KNN trained graders first, for
Each unmarked sample, if the result that two graders are classified to it is consistent, the unmarked sample is that confidence level is high
Sample, take the high unlabeled exemplars of all confidence levels, give a respective label value, add it to tally set XlIn,
Update exemplar collection, then using the marker samples collection after renewal as training set to random woods model FsIt is trained, obtains
New Random Forest model Fs+1;
In the present invention, the exemplar after renewal is concentrated, and both comprising the sample for being originally used for label, is also included and is originally used for
Sample without label, therefore be semi-supervised Random Forest model using the Random Forest model of the tally set training after renewal.
Step 5: optimize semi-supervised Random Forest model:
5a, optimized using definitive operation process (Deterministic Annealing process, DA)
Processing, by introducing a category distribution probability without label dataWhereinAdjust, the part without label data is added in optimization aim, expression formula is as follows for normalization:
For the loss function of whole optimization Random Forest model, sum term Part I
To have label data loss, Section 2For the expectation lost without label data, Section 3It is expressed as the comentropy of no label data distribution;α is expectation weighted value and α ∈ without label data loss
[0,1], value is the temperature variable that 0.5, T is annealing in the present embodiment, and initial value is set as 1, and when T is 0, the formula turns to
Initial Random Forest model loss function;In the model, the loss of exemplar is not only contained, and contains nothing
The predicting list loss of exemplar, and the prediction category without label data is lost as optimization aim, and consider pre-
Survey the comentropy of category probability distribution so that expectation is not only ensure that in optimization, and controls variance, maximises all samples
This interval.
The present invention optimizes by using definitive operation mode to semi-supervised Random Forest model, in the model,
The loss of label data is not only allowed for, is also lost the prediction category without label data as optimization aim, and consider
Predict the comentropy of category probability distribution so that expectation is not only ensure that in optimization, and controls variance, is maximised all
The interval of sample.
5b, the data false segmentation rate initial value for giving a block mold:To control optimization, a suboptimization is carried out
After calculate the outer data false segmentation rate of whole bagWhenWhen, stop optimization, it is believed that have been obtained for optimal random forest
Model.Otherwise next suboptimization is carried out, goes to step 4, renewal annealing temperature Ts+1=0.9Ts, training iterations S=S+1;
Until having obtained optimal Random Forest model Fs+1。
Step 6: build similarity relation figure with Random Forest model:
6a, in Random Forest model, the number of plies of each decision tree is identical, it is assumed that the number of plies t, a pair of data point (xi,
xj) from root node γ, divide by feature layer by layer, most a pair of data point (x of Zhongdaoi,xj) belonging to child node liAnd lj, a pair
Two data point x in data pointiAnd xjThe path of process is expressed as:
QiRepresent data point xiPath, QjRepresent data point xjPath, γ be decision tree root node.For number
Strong point xiBy the decision tree internal node on path;For data point xjBy the decision tree internal node on path.
6b, for each to data point (xi,xj), with the similarity relation represented between them, the similarity relation of all data
A similar diagram matrix W is formed, similarity relation figure matrix W has been obtained, has just obtained similarity relation figure, then had:
wijIt is data point xiWith data point xjBetween similarity relation value.
The present invention is using being suitable for that there is the random forest method of non-linear behavior polarization SAR data to be patterned, therefore
Obtained similarity relation figure matrix W has more preferable compatibility.
Step 7: structure spatial information figure:For each pixel samples point in image, it and surrounding all around four
Individual pixel samples point similarity relation value is 1, and then value is 0 with other pixel samples point similarity relations, so as to obtain a table
Show the figure matrix of spatial relation between whole image sample point, be designated as spatial information figure matrix G.
Step 8: merge similarity relation figure and spatial information figure:By in the similar diagram matrix W and step 7 in step 6
Spatial information figure matrix G, which is combined, draws the final figure matrix Z=W+ for representing similarity relation between pixel samples point in image
λ G, wherein λ ∈ (0,1), take λ=0.5 in the present embodiment.
Step 9: Polarimetric SAR Image is classified and calculates accuracy:
To polarization SAR figure in 9a, the semisupervised classification method for keeping obtained similarity relation figure matrix Z for figure
As being classified, it is as follows to obtain each pixel class label matrix Y, formula in image:
Wherein zijThe element arranged for the i-th row jth in Z, yiFor ith pixel point xiClass label, yjFor j-th of pixel
Point xjClass label, as pixel xiWith pixel xjSimilarity is higher, i.e. zijWhen value is very big, obtain above-mentioned formula
To optimal solution, then yiAnd yjValue it is close, takeThat is ith pixel point xiWith j-th of pixel xjIt is assigned to same
Class, the class label matrix of whole image data is Y=(y after classification1,y2,…yi,…yj…yn);
9b, obtain the sorted class label matrix Y of each pixel;According to class label matrix Y to each pixel
Coloured, each pixel corresponded on Polarimetric SAR Image, using red, green, blueness as three primary colours, according to three bases
Color method is painted for each pixel on color, the result figure after output category, referring to Fig. 2 (f);
9c, the class label matrix Y that grader is predicted and the real class label matrix L of test sample are contrasted,
Draw the classification accuracy rate of experiment.
Method of the invention by using improved semi-supervised random forest, using image pattern point self attributes random
The similarity relation figure matrix between image pixel sample point is instructed to build in the path that decision tree is divided in forest model, such
To similarity relation figure matrix more conform to the nonlinear characteristics of polarization SAR data.Polarization is reasonably make use of on this basis
Spatial information in SAR image sample point, so as to advantageously in the terrain classification of Polarimetric SAR Image.
The technique effect of the present invention is illustrated below by emulation
Embodiment 6
Semi-supervised Classification of Polarimetric SAR Image method based on random forest composition with embodiment 1-5,
Experiment condition
Microcomputer CPU used in experiment is Intel Corei5-2430M internal memory 4GB, and programming platform is Matlab R2011b.
The farmland analogous diagram that experiment figure is 120 × 150, the figure one share 18000 pixels, nine class crops, taken
For wherein 1% pixel as training sample, remaining is test sample.
Experiment content
The present invention utilizes the improved semi-supervised gloomy composition of stochastic model, the semisupervised classification method pair kept then in conjunction with figure
Polarization SAR geo-objects simulation figure is classified, and is carried out on the premise of same Setup Experiments with other Classification of Polarimetric SAR Image methods
Compare, wherein KNN is the sorting technique of supervision, and Fig. 2 (c) is the result figure classified by KNN methods to Fig. 2 (a);Hyper
For the semisupervised classification method based on hypergraph, the result figure that Fig. 2 (d) is classified with Hyper methods to Fig. 2 (a);SSA is base
In the semisupervised classification method of anchor point space diagram, Fig. 2 (e) is the result figure classified using SSA methods to Fig. 2 (a);SSRF
For the inventive method, Fig. 2 (f) is the result figure that the inventive method is classified to Fig. 2 (a).
Table 1 is the terrain classification precision and totality of the polarization SAR geo-objects simulation image respectively obtained using above-mentioned 4 kinds of methods
Nicety of grading
The terrain classification precision and overall classification accuracy of table 1, various methods in analogous diagram
From table 1 it follows that in the case where training sample is 1%, it is of the invention to divide with existing Polarimetric SAR Image
Class method, which is compared, has higher nicety of grading.And the inventive method is very high for the overall classification accuracy rate of image, reaches
99.13%, and worst KNN methods are only 86.03.It is worth noting that, the inventive method is designated as 4 and category for class
Reach 100% for 9 image pattern point classification accuracy rate.
Embodiment 7
Semi-supervised Classification of Polarimetric SAR Image method based on random forest composition is with embodiment 1-5, the condition of emulation and interior
Hold with embodiment 6.
Fig. 2 (a) is the inventive method and the existing original graph for contrasting algorithm and being used for testing, and Fig. 2 (b) is that original graph is corresponding
Standard drawing, Fig. 2 (c) is the result figure classified by KNN methods to Fig. 2 (a), and Fig. 2 (d) is with Hyper methods to Fig. 2 (a)
The result figure classified, Fig. 2 (e) are the result figures classified using SSA methods to Fig. 2 (a), and Fig. 2 (f) is present invention side
The result figure that method is classified to Fig. 2 (a).
The classifying quality of KNN methods is poor it can be seen from Fig. 2 (c), has located the blue region in the lower right corner, other regions
Noise all it has been covered with, overall classification results are very poor, red color area of the Hyper methods to the image upper left corner it can be seen from Fig. 2 (d)
The gray area classifying quality of domain and image right middle section is very poor, and noise is a lot, the SSA methods it can be seen from Fig. 2 (e)
Lifted for the classification accuracy rate in the very poor region of Hyper classifying qualities, but for the yellow area in the upper right corner in figure
Classification, SSA methods are not so good as Hyper, compares figure 2 (e) as can be seen that the inventive method for each piece of region in image point
Class effect is better than other several contrast algorithms, the red color area in the special upper left corner all bad for three contrast algorithm classification effects
Domain, the inventive method also only have only a few noise, what is particularly worth mentioning is that, the inventive method is for powder middle above image
Red area and the blue region in the image lower right corner, classification accuracy rate have nearly reached 100%.
Referring to the simulation experiment result Fig. 2, the relative present invention that demonstrates has higher visuality.
Semi-supervised Classification of Polarimetric SAR Image method proposed by the present invention based on random forest composition avoids existing pole
Change the deficiency that similarity relation between sample point distance metric sample point in image is utilized in SAR image sorting technique, can be effective
The nicety of grading of Polarimetric SAR Image is improved, and higher nicety of grading can be also obtained in the case where training sample is less.
In summary, a kind of semi-supervised Classification of Polarimetric SAR Image method based on random forest composition of disclosure of the invention.Solution
Similarity relation between sample point can not accurately be represented in image in existing Classification of Polarimetric SAR Image method of having determined, fails to fill
Divide the technical problem using spatial information between image pattern point.Its step includes:1st, input picture:From polarization SAR picture number
According to the initial data that Polarimetric SAR Image is obtained in file;2nd, feature extraction:To its progress of the polarization SAR initial data of input
Decompose, extract relevant T features, Cloud features, Freeman features, the Span features of data, obtain the data set of 15 dimensions altogether
X;3rd, initial random forest model is built:Data set X is upset at random, builds property set respectively, wherein first property set is
X1, comprising relevant T features, Cloud features, Span features, second property set is X2, it is special comprising relevant T features, Freeman
Sign;It is F to build initial forest0, 1% sample is chosen as exemplar, by the loss of exemplar and unlabeled exemplars
Loss is combined, and initial random forest F is built using same loss function to unlabeled exemplars and exemplar0;4th, train
Semi-supervised random forest:Initialization training iterations S=0, s exemplar is chosen per class, with first property set X1Instruction
Practice first KNN graders f1, with second property set X2Train second KNN graders f2, and assisted using the two graders
Help random forest to carry out semi-supervised training, unmarked sample is divided by the synergy of two KNN graders first
Classes, the high unlabeled exemplars of confidence level are chosen, their label values is given, is added in tally set, update exemplar collection,
Then using the marker samples collection after renewal as training set to random woods FsHalf is carried out to train;5th, semi-supervised random forest is optimized:
Optimization processing is carried out using definitive operation process (Deterministic Annealing process, DA), by drawing
Enter a category distribution probability without label data, the part without label data is added in optimization aim;Given one whole
The data false segmentation rate initial value of body Model:To control optimization, the outer data mistake of whole bag is calculated after carrying out a suboptimization
Divide rateWhenWhen, stop optimization, otherwise carry out next suboptimization, go to step (4), renewal annealing temperature Ts+1=
0.9Ts, training iterations S=S+1;6th, random forest structure similarity relation figure:Using the random forest trained to image
In exemplar and unlabeled exemplars carry out build similarity relation figure, obtain a similarity relation figure matrix W;7th, space is built
Hum pattern:For each pixel samples point in image, it and the four pixel samples point point similarity relations of surrounding all around
Value is 1, and then value is 0 with other pixel samples point similarity relations, structure spatial information figure matrix G;8th, similarity relation is merged
Figure and spatial information figure:Spatial information figure matrix G in similar diagram matrix W in (6) and (7) is combined draw it is final
Represent figure the matrix Z=W+ λ G, wherein λ ∈ (0,1) of similarity relation between pixel samples point in image;9th, to Polarimetric SAR Image
Classified and calculate classification accuracy rate:The method kept using figure, with reference to obtained similarity relation figure matrix Z to polarization
SAR image is classified, and obtains the sorted class label matrix Y of each pixel;According to class label matrix Y to each picture
Vegetarian refreshments is coloured, the image after output category, and calculates the accuracy of classification.The present invention is calculated using semi-supervised random forest
Method constructs more affine similarity relation figure and considers spatial information, improves the classification accuracy rate of Polarimetric SAR Image.With
In the civil and military such as geological prospecting, rescue and relief work, target identification field.
Claims (3)
- A kind of 1. Polarimetric SAR Image semisupervised classification method based on random forest composition, it is characterised in that include following step Suddenly:(1) input picture:The initial data of Polarimetric SAR Image is obtained from polarimetric SAR image data;(2) feature extraction:To the polarization SAR initial data of input, it is decomposed, and relevant T features, the Cloud for extracting data are special Sign, Freeman features, Span features, obtain the data set X of 15 dimensions altogether;(3) initial random forest model is built:Data set X is upset at random, builds property set respectively, wherein first property set For X1, comprising relevant T features, Cloud features, Span features, second property set is X2, it is special comprising relevant T features, Freeman Sign;It is F to build initial forest0, the sample per class 1% is chosen as exemplar collection, is designated as Xl, remaining is unlabeled exemplars Collection, is designated as Xu, the loss of exemplar and the loss of unlabeled exemplars are combined, to unlabeled exemplars and exemplar using same One loss function builds initial random forest model F0;(4) semi-supervised Random Forest model is trained:Initialization training iterations S=0, s label sample is chosen from exemplar Originally two graders are respectively trained, with first property set X1In s exemplar train first KNN graders f1, with Two property set X2In s exemplar train second KNN graders f2, for each unmarked sample, if two points The result that class device is classified to it is consistent, then the unmarked sample is the high sample of confidence level, takes all confidence levels high Unlabeled exemplars, a respective label value is given, adds it to tally set XlIn, exemplar collection is updated, then will renewal Marker samples collection afterwards is as training set to random woods model FsSemi-supervised training is carried out, obtains new Random Forest model Fs+1;(5) semi-supervised Random Forest model is optimized:Optimization processing is carried out using definitive operation process, by introducing a nothing The category distribution probability of label dataData in all unlabeled exemplars are added in optimization aim, give an entirety The data false segmentation rate initial value of model:To control optimization, the outer data of whole bag are calculated after one suboptimization of progress and are divided by mistake RateWhenWhen, stop optimization, it is believed that have been obtained for optimal Random Forest model.Otherwise next suboptimum is carried out Change, go to step (4), renewal annealing temperature Ts+1=0.9Ts, training iterations S=S+1;Until obtain optimal random Forest model;(6) similarity relation figure is built with Random Forest model:Using the Random Forest model trained to the label sample in image This and unlabeled exemplars build similarity relation figure, obtain a similarity relation figure matrix W;(7) spatial information figure is built:For each pixel samples point in image, it and the four pixel samples of surrounding all around This point point similarity relation value is 1, and then value is 0 with other pixel samples point similarity relations, structure spatial information figure matrix G;(8) similarity relation figure and spatial information figure are merged:Similar diagram matrix W and spatial information figure matrix G are combined and drawn most Similarity relation figure matrix Z=W+ λ G, wherein λ ∈ (0,1) between pixel samples point in whole expression image;(9) Polarimetric SAR Image is classified and calculates classification accuracy rate:The method kept using figure, with obtained phase Polarimetric SAR Image is classified like graph of a relation matrix Z, obtains the sorted class label matrix Y of each pixel;According to class Distinguishing label matrix Y colours to each pixel, the image after output category, completes Polarimetric SAR Image semisupervised classification, and Calculate classification accuracy rate.
- 2. the Polarimetric SAR Image semisupervised classification method based on random forest composition according to claims 1, wherein walking Suddenly definitive operation optimizing expression is in (5)<mrow> <msub> <mi>L</mi> <mrow> <mi>D</mi> <mi>A</mi> </mrow> </msub> <mrow> <mo>(</mo> <mrow> <mi>g</mi> <mo>,</mo> <mover> <mi>p</mi> <mo>^</mo> </mover> </mrow> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <mo>|</mo> <msub> <mi>X</mi> <mi>l</mi> </msub> <mo>|</mo> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mrow> <mi>x</mi> <mo>&Element;</mo> <msub> <mi>X</mi> <mi>l</mi> </msub> </mrow> </munder> <msub> <mi>l</mi> <mi>l</mi> </msub> <mrow> <mo>(</mo> <mrow> <mi>g</mi> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> <mo>)</mo> </mrow> <mo>+</mo> <mfrac> <mi>&alpha;</mi> <mrow> <mo>|</mo> <msub> <mi>X</mi> <mi>u</mi> </msub> <mo>|</mo> </mrow> </mfrac> <munder> <mi>&Sigma;</mi> <mrow> <mi>x</mi> <mo>&Element;</mo> <msub> <mi>X</mi> <mi>u</mi> </msub> </mrow> </munder> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>k</mi> </munderover> <mi>p</mi> <mrow> <mo>(</mo> <mrow> <mover> <mrow> <mi>i</mi> <mo>|</mo> </mrow> <mo>^</mo> </mover> <mi>x</mi> </mrow> <mo>)</mo> </mrow> <msub> <mi>l</mi> <mi>u</mi> </msub> <mrow> <mo>(</mo> <mrow> <msub> <mi>g</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> </mrow> <mo>)</mo> </mrow> <mo>+</mo> <mfrac> <mi>T</mi> <msub> <mi>X</mi> <mi>u</mi> </msub> </mfrac> <munder> <mi>&Sigma;</mi> <mrow> <mi>x</mi> <mo>&Element;</mo> <msub> <mi>X</mi> <mi>u</mi> </msub> </mrow> </munder> <munderover> <mi>&Sigma;</mi> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>k</mi> </munderover> <mi>H</mi> <mrow> <mo>(</mo> <mover> <mi>p</mi> <mo>^</mo> </mover> <mo>)</mo> </mrow> </mrow>Sum term Part ITo have label data loss, Section 2For Expectation without label data loss, Section 3It is expressed as the comentropy of no label data distribution;α is without mark The expectation weighted value and α ∈ [0,1] of data degradation are signed, value be that 0.5, T is the temperature variable annealed in the present embodiment, can be with Find out, when T is 0, the formula turns to model in (3);In the model, the loss of exemplar is contained, contains no mark The predicting list loss of signed-off sample sheet, and the prediction category without label data is lost as optimization aim, maximise all The interval of sample.
- 3. the Polarimetric SAR Image semisupervised classification method according to claim 1 based on random forest composition, its feature exist In, described in step (6) with Random Forest model build similarity relation figure, specifically included:In Random Forest model, the number of plies of each decision tree is identical, it is assumed that the number of plies t, a pair of data point (xi,xj) from root Node γ, divide by feature layer by layer, most a pair of data point (x of Zhongdaoi,xj) belonging to child node liAnd lj, a pair of data points In two data point xiAnd xjThe path of process is expressed as:<mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msup> <mi>Q</mi> <mi>i</mi> </msup> <mo>=</mo> <mo>{</mo> <mi>&gamma;</mi> <mo>,</mo> <msubsup> <mi>s</mi> <mn>1</mn> <mi>i</mi> </msubsup> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msubsup> <mi>s</mi> <mi>k</mi> <mi>i</mi> </msubsup> <mo>}</mo> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msup> <mi>Q</mi> <mi>j</mi> </msup> <mo>=</mo> <mo>{</mo> <mi>&gamma;</mi> <mo>,</mo> <msubsup> <mi>s</mi> <mn>1</mn> <mi>j</mi> </msubsup> <mo>,</mo> <mo>...</mo> <mo>,</mo> <msubsup> <mi>s</mi> <mn>2</mn> <mi>j</mi> </msubsup> <mo>}</mo> </mrow> </mtd> </mtr> </mtable> </mfenced>The similarity relation of all data forms a similar diagram matrix W, then has:<mfenced open = "{" close = ""> <mtable> <mtr> <mtd> <mrow> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mn>1</mn> </mrow> </mtd> <mtd> <mrow> <mi>i</mi> <mi>f</mi> <mi> </mi> <msup> <mi>Q</mi> <mi>i</mi> </msup> <mo>=</mo> <msup> <mi>Q</mi> <mi>j</mi> </msup> </mrow> </mtd> </mtr> <mtr> <mtd> <mrow> <msub> <mi>w</mi> <mrow> <mi>i</mi> <mi>j</mi> </mrow> </msub> <mo>=</mo> <mn>0</mn> </mrow> </mtd> <mtd> <mrow> <mi>o</mi> <mi>t</mi> <mi>h</mi> <mi>e</mi> <mi>r</mi> <mi>e</mi> <mi>e</mi> <mi>l</mi> <mi>s</mi> <mi>e</mi> </mrow> </mtd> </mtr> </mtable> </mfenced>wijRepresent data point xiWith data point xjBetween similarity relation.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710341085.3A CN107358142B (en) | 2017-05-15 | 2017-05-15 | Polarized SAR image semi-supervised classification method based on random forest composition |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201710341085.3A CN107358142B (en) | 2017-05-15 | 2017-05-15 | Polarized SAR image semi-supervised classification method based on random forest composition |
Publications (2)
Publication Number | Publication Date |
---|---|
CN107358142A true CN107358142A (en) | 2017-11-17 |
CN107358142B CN107358142B (en) | 2020-12-08 |
Family
ID=60272053
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201710341085.3A Active CN107358142B (en) | 2017-05-15 | 2017-05-15 | Polarized SAR image semi-supervised classification method based on random forest composition |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN107358142B (en) |
Cited By (16)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107590507A (en) * | 2017-08-23 | 2018-01-16 | 西安电子科技大学 | The semi-supervised Classification of Polarimetric SAR Image method of quick renewal figure based on space constraint |
CN108446652A (en) * | 2018-03-27 | 2018-08-24 | 武汉大学 | Polarimetric SAR image terrain classification method based on dynamic texture feature |
CN108509935A (en) * | 2018-04-12 | 2018-09-07 | 电子科技大学 | A kind of radar operation mode recognition methods based on random forests algorithm |
CN108629373A (en) * | 2018-05-07 | 2018-10-09 | 苏州大学 | A kind of image classification method, system, equipment and computer readable storage medium |
CN108734228A (en) * | 2018-06-14 | 2018-11-02 | 中交第二公路勘察设计研究院有限公司 | The polarimetric SAR image random forest classification method of comprehensive multiple features |
CN108764310A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | SAR target identification methods based on multiple dimensioned multiple features depth forest |
CN109063760A (en) * | 2018-07-22 | 2018-12-21 | 西安电子科技大学 | Polarization SAR classification method based on the multiple dimensioned convolution model of random forest |
CN109697469A (en) * | 2018-12-26 | 2019-04-30 | 西北工业大学 | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint |
CN110414566A (en) * | 2019-07-01 | 2019-11-05 | 武汉大学 | A kind of Land cover types change detecting method based on time series PolSAR image |
CN110738257A (en) * | 2019-10-15 | 2020-01-31 | 西安电子科技大学 | Optical remote sensing image water area classification method based on feature extraction |
CN111142119A (en) * | 2020-01-10 | 2020-05-12 | 中国地质大学(北京) | Mine geological disaster dynamic identification and monitoring method based on multi-source remote sensing data |
CN111310840A (en) * | 2020-02-24 | 2020-06-19 | 北京百度网讯科技有限公司 | Data fusion processing method, device, equipment and storage medium |
CN112308151A (en) * | 2020-11-03 | 2021-02-02 | 西安电子科技大学 | Weighting-based classification method for hyperspectral images of rotating forest |
CN112712094A (en) * | 2019-10-24 | 2021-04-27 | 北京四维图新科技股份有限公司 | Model training method, device, equipment and storage medium |
CN112749841A (en) * | 2020-12-30 | 2021-05-04 | 科大国创云网科技有限公司 | User public praise prediction method and system based on self-training learning |
CN113792809A (en) * | 2021-09-18 | 2021-12-14 | 中国空间技术研究院 | Remote sensing picture classification method based on random semi-supervised feature extraction model |
Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130346346A1 (en) * | 2012-06-21 | 2013-12-26 | Microsoft Corporation | Semi-supervised random decision forests for machine learning |
CN105528595A (en) * | 2016-02-01 | 2016-04-27 | 成都通甲优博科技有限责任公司 | Method for identifying and positioning power transmission line insulators in unmanned aerial vehicle aerial images |
CN106056134A (en) * | 2016-05-20 | 2016-10-26 | 重庆大学 | Semi-supervised random forests classification method based on Spark |
CN106096627A (en) * | 2016-05-31 | 2016-11-09 | 河海大学 | The Polarimetric SAR Image semisupervised classification method that considering feature optimizes |
-
2017
- 2017-05-15 CN CN201710341085.3A patent/CN107358142B/en active Active
Patent Citations (4)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US20130346346A1 (en) * | 2012-06-21 | 2013-12-26 | Microsoft Corporation | Semi-supervised random decision forests for machine learning |
CN105528595A (en) * | 2016-02-01 | 2016-04-27 | 成都通甲优博科技有限责任公司 | Method for identifying and positioning power transmission line insulators in unmanned aerial vehicle aerial images |
CN106056134A (en) * | 2016-05-20 | 2016-10-26 | 重庆大学 | Semi-supervised random forests classification method based on Spark |
CN106096627A (en) * | 2016-05-31 | 2016-11-09 | 河海大学 | The Polarimetric SAR Image semisupervised classification method that considering feature optimizes |
Non-Patent Citations (1)
Title |
---|
CUI XIE ET AL: "Object-Oriented Random Forest Classification for Enteromorpha Prolifera", 《2016 INTERNATIONAL CONFERENCE ON VIRTUAL REALITY AND VISUALIZATION》 * |
Cited By (25)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN107590507B (en) * | 2017-08-23 | 2020-06-09 | 西安电子科技大学 | Semi-supervised polarimetric SAR image classification method for fast updating map based on space constraint |
CN107590507A (en) * | 2017-08-23 | 2018-01-16 | 西安电子科技大学 | The semi-supervised Classification of Polarimetric SAR Image method of quick renewal figure based on space constraint |
CN108446652A (en) * | 2018-03-27 | 2018-08-24 | 武汉大学 | Polarimetric SAR image terrain classification method based on dynamic texture feature |
CN108446652B (en) * | 2018-03-27 | 2021-09-03 | 武汉大学 | Polarized SAR image ground feature classification method based on dynamic texture features |
CN108509935A (en) * | 2018-04-12 | 2018-09-07 | 电子科技大学 | A kind of radar operation mode recognition methods based on random forests algorithm |
CN108509935B (en) * | 2018-04-12 | 2020-01-03 | 电子科技大学 | Radar working mode identification method based on random forest algorithm |
CN108629373A (en) * | 2018-05-07 | 2018-10-09 | 苏州大学 | A kind of image classification method, system, equipment and computer readable storage medium |
CN108764310A (en) * | 2018-05-17 | 2018-11-06 | 西安电子科技大学 | SAR target identification methods based on multiple dimensioned multiple features depth forest |
CN108764310B (en) * | 2018-05-17 | 2021-10-29 | 西安电子科技大学 | SAR target recognition method based on multi-scale multi-feature depth forest |
CN108734228A (en) * | 2018-06-14 | 2018-11-02 | 中交第二公路勘察设计研究院有限公司 | The polarimetric SAR image random forest classification method of comprehensive multiple features |
CN109063760A (en) * | 2018-07-22 | 2018-12-21 | 西安电子科技大学 | Polarization SAR classification method based on the multiple dimensioned convolution model of random forest |
CN109063760B (en) * | 2018-07-22 | 2021-06-08 | 西安电子科技大学 | Polarization SAR classification method based on random forest multi-scale convolution model |
CN109697469A (en) * | 2018-12-26 | 2019-04-30 | 西北工业大学 | A kind of self study small sample Classifying Method in Remote Sensing Image based on consistency constraint |
CN110414566A (en) * | 2019-07-01 | 2019-11-05 | 武汉大学 | A kind of Land cover types change detecting method based on time series PolSAR image |
CN110414566B (en) * | 2019-07-01 | 2020-11-17 | 武汉大学 | Land coverage type change detection method based on time sequence PolSAR image |
CN110738257A (en) * | 2019-10-15 | 2020-01-31 | 西安电子科技大学 | Optical remote sensing image water area classification method based on feature extraction |
CN112712094A (en) * | 2019-10-24 | 2021-04-27 | 北京四维图新科技股份有限公司 | Model training method, device, equipment and storage medium |
CN111142119B (en) * | 2020-01-10 | 2021-08-17 | 中国地质大学(北京) | Mine geological disaster dynamic identification and monitoring method based on multi-source remote sensing data |
CN111142119A (en) * | 2020-01-10 | 2020-05-12 | 中国地质大学(北京) | Mine geological disaster dynamic identification and monitoring method based on multi-source remote sensing data |
CN111310840A (en) * | 2020-02-24 | 2020-06-19 | 北京百度网讯科技有限公司 | Data fusion processing method, device, equipment and storage medium |
CN111310840B (en) * | 2020-02-24 | 2023-10-17 | 北京百度网讯科技有限公司 | Data fusion processing method, device, equipment and storage medium |
CN112308151A (en) * | 2020-11-03 | 2021-02-02 | 西安电子科技大学 | Weighting-based classification method for hyperspectral images of rotating forest |
CN112749841A (en) * | 2020-12-30 | 2021-05-04 | 科大国创云网科技有限公司 | User public praise prediction method and system based on self-training learning |
CN113792809A (en) * | 2021-09-18 | 2021-12-14 | 中国空间技术研究院 | Remote sensing picture classification method based on random semi-supervised feature extraction model |
CN113792809B (en) * | 2021-09-18 | 2024-07-02 | 中国空间技术研究院 | Remote sensing picture classification method based on random semi-supervised feature extraction model |
Also Published As
Publication number | Publication date |
---|---|
CN107358142B (en) | 2020-12-08 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN107358142A (en) | Polarimetric SAR Image semisupervised classification method based on random forest composition | |
CN105303198B (en) | A kind of remote sensing image semisupervised classification method learnt from fixed step size | |
CN104077599B (en) | Polarization SAR image classification method based on deep neural network | |
Xu et al. | Decision tree regression for soft classification of remote sensing data | |
Mas | Mapping land use/cover in a tropical coastal area using satellite sensor data, GIS and artificial neural networks | |
CN104408466B (en) | Learn the high-spectrum remote sensing semisupervised classification method of composition based on local manifolds | |
CN108846426A (en) | Polarization SAR classification method based on the twin network of the two-way LSTM of depth | |
CN107145836B (en) | Hyperspectral image classification method based on stacked boundary identification self-encoder | |
CN109614973A (en) | Rice seedling and Weeds at seedling image, semantic dividing method, system, equipment and medium | |
CN111709318B (en) | High-resolution remote sensing image classification method based on generation countermeasure network | |
CN112070079B (en) | X-ray contraband package detection method and device based on feature map weighting | |
CN105608690A (en) | Graph theory and semi supervised learning combination-based image segmentation method | |
CN108492298A (en) | Based on the multispectral image change detecting method for generating confrontation network | |
CN104298974A (en) | Human body behavior recognition method based on depth video sequence | |
CN106683102A (en) | SAR image segmentation method based on ridgelet filters and convolution structure model | |
CN104182767A (en) | Active learning and neighborhood information combined hyperspectral image classification method | |
Yao et al. | Sensing urban land-use patterns by integrating Google Tensorflow and scene-classification models | |
Durduran | Automatic classification of high resolution land cover using a new data weighting procedure: The combination of k-means clustering algorithm and central tendency measures (KMC–CTM) | |
Xu et al. | Multi-modal deep learning for weeds detection in wheat field based on RGB-D images | |
CN105809119A (en) | Sparse low-rank structure based multi-task learning behavior identification method | |
Li et al. | Multi-scale object detection in satellite imagery based on YOLT | |
CN112084860A (en) | Target object detection method and device and thermal power plant detection method and device | |
Du et al. | Training SegNet for cropland classification of high resolution remote sensing images | |
CN107437091A (en) | Multilayer limits the positive and negative class change detecting method of SAR image of Boltzmann machine | |
Khalid et al. | Evaluation the accuracy of oil palm tree detection using deep learning and support vector machine classifiers |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination | ||
GR01 | Patent grant | ||
GR01 | Patent grant |